This is general education for business owners and executives, not financial or accounting advice for your specific company. Your CFO and your accountant should sign off on the real numbers.
Almost every executive who calls Torque AI asks some version of the same question in the first ten minutes: what is this going to cost us? It is the right question, and most of the answers in the market are either a vendor's seat price, which is too low, or a consultancy's enterprise estimate, which is too high. Here is how I break the budget down for a mid-market company.
Where does the money in an AI rollout actually go?
Four buckets, in rising order of size for a typical first project.
- Software. The AI tools themselves, priced by the seat or by usage.
- Integration. Connecting those tools to your CRM, practice management or ERP system, your phone system and your documents, and making sure the data flowing into them is clean enough to trust.
- People time. Your team's hours mapping the process, testing, correcting and changing how they work.
- Governance and training. The usage policy, the review steps, the vendor checks and the hours spent teaching staff what the system may and may not do.
Most published budgets count the first bucket carefully, the second bucket roughly, and the last two not at all. That is why so many AI projects are described as "over budget" when they were simply never fully budgeted.
What does the software itself cost?
Less than people expect. General assistants for staff are priced per seat per month at roughly the cost of a mid-tier CRM or accounting license. Voice and chat agents that answer calls and forms are priced by minutes or conversations and, for most companies, come to a few hundred to a couple of thousand dollars a month. Usage-based model access for a custom system is often the smallest line of all until volume gets large.
The useful comparison is one salary. A first project's software for a year typically costs a fraction of one entry-level hire. If a vendor's quote is several times that for a first deployment, you are being sold an enterprise contract, and you should ask what a 90-day pilot costs instead.
Why is integration the line item that surprises people?
Because the model is the easy part. The work is getting the model to see your data and write back to your systems. A chat agent that cannot create a lead record in your CRM is a toy. A drafting assistant that cannot read your templates and past files is a web page. Integration is where a project becomes a system, and it is also where the hours pile up: field mapping, permissions, cleaning up years of inconsistent records, and testing every path a customer or a document can take.
For a single system connected to one or two platforms, integration is usually a one-time cost in the low five figures when done by an outside team, less if your platforms have native connectors and someone on staff can own them. Budget for a second pass 60 days in. The first version is never quite right, and the fixes are cheap if you planned for them and expensive if you did not.
How do you budget for people time?
Write it down as a real number. Take the three or four people who will map the process, test the system and change their daily routine, estimate their hours over 90 days, and multiply by their loaded hourly cost. For most first projects this comes to the same order of magnitude as software and integration combined, which is why I use the rule of one internal dollar for every external dollar in year one.
Two things make this number smaller. First, pick a project with a narrow scope and a clear owner, the way I described for law firms adopting AI for client acquisition: one intake path, one document type, one follow-up sequence. Second, run the new system alongside the old one for two weeks rather than switching over on a Monday. The parallel period costs hours, but it costs far fewer hours than a failed cutover.
What should you budget for governance and training?
A few days of focused work and a small recurring commitment, which makes it the best value on the sheet. A one-page AI usage policy, a short list of what the system may answer and what it must hand to a person, a vendor review covering where your data goes, and a two-hour training session for everyone who touches the system. Then a monthly review of logs or transcripts. For a mid-market company this is measured in hours, not months.
Skipping it is the expensive option. A system with no boundary produces one bad customer answer, one leaked document or one unverified payment change, and the cleanup costs more than the whole governance budget. The point is not to slow the project down. It is to make the rest of the budget safe to spend.
How do you build the business case before you spend anything?
Price the problem first. Count the hours the project will remove, the leads that currently go unanswered, or the errors it will catch, and put a dollar figure on each. Then run the exercise I call premortem-first: write the memo that explains why the project failed, dated a year from now, and fix the top three causes before you sign anything. If the fixed problem is worth less than about three times the all-in first-year cost, choose a different project. There is always one that clears the bar.
When I helped a five-person team generate $32 million in revenue, we did not have a large technology budget. We had a tight process for the moments that made money, and we automated those first. That is still the rule. Spend on the part of the revenue engine that is leaking, not on the part that is interesting.
What does a sample 90-day budget look like?
Sample first-project budget, mid-market company, 90 days
Software: seats or usage for the one system you are deploying, three months. Expect a few hundred to a few thousand dollars a month.
Integration: one-time connection to one or two core platforms, plus a second pass at day 60. Typically low five figures with outside help, less with native connectors and an internal owner.
People time: hours for the process owner, two testers and the manager who approves changes, at loaded cost. Plan for roughly the same total as software plus integration.
Governance and training: usage policy, vendor review, boundary list, one training session, monthly log review. A few days of work, repeated lightly each month.
Contingency: 15 to 20 percent of the total. Every first project finds a data problem nobody knew about.
What to measure: hours saved per week, speed to first contact, quotes or proposals started, error rate, and adoption, meaning how many of the people who should use it actually do.
If you want the numbers for your company rather than a general range, the AI Clarity Sprint produces three ROI-modeled recommendations and a 90-day plan in three weeks, at a fixed fee agreed before we start. Details are on the engagements and FAQ page. For the prompts and procedures behind this kind of work, see ChatGPT for Profit, Volumes 1 and 2.
Frequently asked questions
- How much does an AI rollout cost a mid-market company?
For a company with 50 to 500 employees, a first focused deployment usually lands in the tens of thousands of dollars over the first 90 days once software, integration, staff time and governance are all counted. Ongoing software is typically a few hundred to a few thousand dollars a month. The seven and eight figure numbers in the press describe enterprise programs, not a first project.
- What is the biggest cost in an AI rollout?
People time. Mapping the process, cleaning the data, testing the system and changing how the team works usually costs more than the software, and it is the line most companies leave off the budget. Plan for roughly one dollar of internal time for every dollar of software and integration in the first year.
- Why is the software the smallest line?
Most business AI tools are sold by the seat or by usage and cost about what a mid-tier CRM or accounting seat costs. The expense is not the model. It is connecting the model to your data and systems and getting the team to use it the same way every day.
- How do you know whether an AI project will pay for itself?
Price the problem before you price the tool. Count the hours, the lost leads or the error rate the project is meant to fix, put a dollar figure on them, and run a premortem on how the project could fail. If the fixed problem is worth less than three times the all-in first-year cost, pick a different project.
- Should a mid-market company hire an AI team or use outside help?
Start with a fixed-scope outside engagement to pick the first project, build it and write the governance rules, and name one internal owner from day one. Hire a full-time AI role only once there are two or three systems running that need daily attention. Most companies under 500 employees do not need a dedicated team in year one.
Want a real budget instead of a range?
The AI Clarity Sprint gives mid-market companies, law firms, wealth managers and insurance agencies an AI audit, deployment recommendations with ROI models, and a 90-day plan in three weeks.